| Legal documents are documents used by judicial organs and lawyers when solving cases,and contain a large amount of judicial information.How to efficiently analyze and use this information is of great significance to research.Most of the existing intelligent processing of legal documents focuses on legal entity recognition and legal entity relationship extraction,and there is very little research on event extraction.As the most common type of documents in daily life,civil judgments are universal to a certain extent.At the same time,judgments are the results of judgments,and they can be used as data sources to obtain detailed descriptions of the details of the cases.In this thesis,civil judgments are downloaded from the magic weapon of Peking University as the original corpus,and sentence-level event extraction is carried out for civil judgments.Event extraction is mainly divided into three tasks: event type identification,trigger word acquisition and event meta identification.For event extraction,the thesis mainly carries out the following four aspects:(1)Collection and preprocessing of civil judgments.This article downloads eight types of documents,including contract disputes,marriage disputes,traffic accident disputes,property rights disputes,personality rights disputes,inheritance disputes,property disputes,and property disputes from the magic weapon of Peking University.After preprocessing,such as expression processing,word segmentation,and removing stop words in the legal field,a total of 11,277 sentences containing the court’s judgment results were obtained.(2)Event type clustering.Because the existing laws and regulations do not have a unified definition for the types of legal events,and their types are diverse,and clustering is an unsupervised learning method,which can automatically cluster similar data.Clustering is used to obtain event types.In this thesis,the K-means clustering method is used for clustering,and the clustering effectiveness measure index silhouette coefficient is used to determine the optimal number of clusters.The experiment shows that the silhouette coefficient is the highest when the clustering is 8 types,and the clustering accuracy rate reaches 82.12.%.(3)Trigger word recognition.This thesis designs a combination of Text Rank(text ranking)algorithm and information entropy to obtain the trigger word set in the sentence set,which is basically consistent with the trigger word set obtained by manual annotation.The clustered event type name is displayed.(4)For the sentence sets for which event types and trigger words have been obtained,this thesis defines the event elements corresponding to different types of events.The relevant event elements are manually marked,and the next event element identification task is carried out.The data obtained in the previous step is divided into training set and test set according to8:2 for model training and testing.This thesis conducts event meta-recognition experiments through a model combining ALBERT(lightweight pre-training model)and BiGRU-Attention-CRF(bidirectional gated recurrent unit-attention mechanism-conditional random field),and results with other recognition models In comparison,the accuracy of this model reaches 83.39%,which is higher than other event meta-recognition models.The experimental results show that the method in the thesis can effectively complete the event type identification,trigger word acquisition and event meta identification.It provides a way of thinking for the research on event extraction of civil judgments,and can provide help for subsequent case review,intelligent legal robots and other applications at other levels. |
